RDP 2007-03: Forecasting with Factors: The Accuracy of Timeliness 1. Introduction

Dynamic factor models, which enable a large number of economic time series to be combined, have been shown to frequently produce more accurate forecasts than standard time-series benchmarks, structural models and even official forecasts, especially at horizons of between one and two years.[1] There are good reasons for this. Using a large number of series can produce a measure of current economic activity that contains less noise than the individual series used in more traditional forecasting frameworks. Further, using many series will typically make the forecasts more robust to structural change in individual explanatory variables and possibly to structural change among the relationships between economic variables. However, there are potential costs to using many series when forecasting. One which has received little attention in the literature is that a larger panel of series will contain less timely series and so can only be used to produce forecasts with a considerable delay. In this paper we examine how forecast accuracy changes when factor forecasts are made using a more timely panel which necessarily contains less information. This is an important trade-off for policy-makers: more accurate forecasts will help policy-makers, but so too will more timely forecasts.

In this paper we examine this trade-off between accuracy and timeliness for eight key Australian macroeconomic series. More timely factor forecasts can be made by using a narrower panel that effectively excludes the information in series with late release dates. Using a narrower panel is likely to reduce the precision of the estimates of the factors, but it remains an open question whether the resulting deterioration in forecast accuracy is large. Several papers have found that the deterioration in forecast accuracy when using smaller panels is not large. In this regard, it is relevant that Boivin and Ng (2006) and Watson (2003) find that forecast accuracy does not improve beyond the use of 50–100 series. Similarly, Schneider and Spitzer (2004) found that they needed to restrict the size of the panel to outperform simple benchmarks when forecasting Austrian GDP.

There are, however, other limitations to the use of factor forecasts for policy purposes that we do not address in this paper. One of the most significant of these is that they are reduced-form forecasts based only on contemporaneous information and so they cannot be conditioned on particular assumptions. Notably, the forecasts cannot be conditioned on a specific path of interest rates. This may limit their usefulness in a policy environment, such as a central bank.

The remainder of this paper proceeds as follows. In Section 2 we briefly describe the factor forecasting procedure. We discuss the timeliness of data and the composition of the panel in Section 3. In Section 4 we document the performance of factor forecasts and the empirical trade-off between accuracy and timeliness, before concluding in Section 5.


This finding has most frequently been demonstrated for inflation: see Angelini, Henry and Mestre (2001) for the euro area; Stock and Watson (1999), Brave and Fisher (2004) and Gavin and Kliesen (2006) for the United States; Gosselin and Tkacz (2001) for Canada; and Moser, Rumler and Scharler (2004) for Austria. Other studies suggest that these results generalise to real variables and interest rates: see Banerjee, Marcellino and Masten (2005) for five new European Union member states; Matheson (2005) for New Zealand; Stock and Watson (2002b) for the United States; and Artis, Banerjee and Marcellino (2005) for the United Kingdom. [1]